<p>Precise prediction of the UBC of shallow foundations subjected to vertical, horizontal, and moment loads in stratified soil is a major concern in geotechnical engineering. Traditional analytical methods may be constrained by several unrealistic assumptions. On the other hand, modern numerical models such as Finite Element Limit Analysis (FELA), although accurate, are computationally expensive to use in practical scenarios. This work presents a new approach, namely PINNs, which incorporates fundamental principles of soil mechanics directly into the machine learning model to predict the UBC rapidly and efficiently. A complete database made up of 10,000 accurate FELA simulations was created for sand-over-clay and clay-over-sand soil strata under V-H-M loading. The designed PINN utilized mechanical equilibrium laws and Mohr–Coulomb failure law in the loss function to ensure mechanical consistency while training. The designed model showed a very good ability in predicting the response of FELA with R² equal to 0.992 and MAPE equal to 2.3%, which is a great improvement compared to artificial neural networks (ANN). Furthermore, the designed framework was able to capture complex failure modes such as punch-through failures, thin-layer effects, and switching between punch failure and general shear failure. Notably, the PINN model cut down computation time for predicting one FELA simulation from about 180&#xa0;s to about 0.002&#xa0;s. The PINN framework suggests an effective model that can predict the ultimate bearing capacity of layered soils with precision. The use of physics-constrained data-driven modeling, with incorporation of the physics of the underlying process, makes the model robust, stable, and generalizable for geotechnical design problems with limited data. PINN holds immense promise for future applications of the technique in the areas of geotechnical engineering, real-time design of foundations, probabilistic geotechnical analysis, and reliability-based geotechnical engineering practice.</p>